Professional Documents
Culture Documents
sciences
Article
Functional Electrical Stimulation Controlled by Motor
Imagery Brain-Computer Interface for Rehabilitation
Inchul Choi 1 , Gyu Hyun Kwon 2 , Sangwon Lee 3 and Chang S. Nam 1, *
1 Fitts Department of Industrial and Systems Engineering, North Carolina State University,
Raleigh, NC 27695, USA; ichoi3@ncsu.edu
2 Graduate School of Technology & Innovation Management, Hanyang University, Seoul 04763, Korea;
ghkwon@hanyang.ac.kr
3 Department of Interaction Science, Sungkyunkwan University, Seoul 03063, Korea; upcircle@skku.edu
* Correspondence: csnam@ncsu.edu; Tel.: +1-919-515-8140; Fax: +1-919-515-5281
Received: 23 June 2020; Accepted: 31 July 2020; Published: 2 August 2020
1. Introduction
Healthy individuals whose brains and neuromuscular systems enable normal motor functions
can naturally perform Activities of Daily Living (ADLs). Nonetheless, for some people who have
disabilities in these functions due to injury or disease, simple tasks become very difficult or impossible
to do. To assist this population, researchers in many fields, from physical therapy to engineering,
have developed various rehabilitation technologies that help them perform ADLs [1,2]. One such
technology, Functional Electrical Stimulation (FES), delivers electrical impulses to either paralyzed or
impaired limbs to generate artificial muscle contraction [3,4]. In this way, FES helps disabled people
perform ADLs such as walking, reaching, and grasping [5,6]. Some FES devices are controlled by
brain–computer interfaces (BCIs), sometimes called brain–machine interfaces.
In general, BCIs can help people communicate and control devices and applications without
using peripheral nerves and muscle pathways [7]. BCIs are also a potential method to promote
the independence of physically disabled people by means of the BCI’s ability to bypass non-functional
neural pathways [8]. A sensorimotor rhythm (SMR)-based BCI-controlled FES system is a novel
technology that combines the advantages of FES and BCI systems, and allows severely disabled
patients to restore motor functions through the FES system by translating voluntary Motor Imagery
(MI) to physical action [9]. There are many potential benefits of combining SMR-based BCIs and FES
systems, such as the promotion of neuroplasticity [10], the restoration of motor functions by using
voluntary motor intentions [9,11], and providing proprioceptive sensory feedback as a result of their
intentions [12].
Although SMR-based BCI-controlled FES methods seem promising, current studies still have
central issues: (1) ambiguous instruction of MI tasks during training under SMR-based BCI systems, and
(2) difficulties in classifying voluntary MI-evoked SMRs and FES-driven passive-movement-evoked
SMRs when FES is activated. Moreover, (3) only a few studies have examined the feasibility of
classifying two different MI tasks of a single hand, such as grasping and opening, and (4) few studies
have examined human factors and ergonomics (HF/E) perspectives such as subjective mental workload
and user satisfaction in the use of SMR-based BCI-controlled FES systems. This research that is
composed of two phases was conducted to address these issues by developing a new SMR-based BCI
system with visual guidance during training to classify a 2-class MI task in a single hand, as well as
voluntary and passive SMRs (Phase 1), and evaluating the feasibility of the proposed BCI-controlled
FES system by performing sequential goal-oriented tasks with stroke and TBI patients (Phase 2).
The remainder of this article consists of five more sections (this introduction being Section 1):
Section 2 describes a survey of current SMR-based BCI studies for FES systems to identify the limitations
of current research and clarifies the current state of BCI-controlled FES technologies. Section 3
presents Phase 1, where an SMR-based BCI system to control FES was developed and validated
to address the issues on current research studies. Section 4 describes Phase 2, which assessed
the feasibility of the proposed BCI-FES system by conducting a sequential task with fixed order
under a semi-asynchronous mode. Section 5 discusses the findings of the present research along with
implications and future directions.
2. Background
then added the term Event-Related Synchronization (ERS) to explain event-related enhancement [53].
These SMR features, ERD and ERS, have been widely employed to decode different MIs, such as left
or right-hand motor intention [54]. SMR-based BCIs allow users who have severe motor disability
(e.g., the paralyzed) to control physical or virtual devices by decoding their motor intention or motor
imagery (e.g., squeezing the left hand or the right hand to move a mouse cursor on the computer
screen to the left or the right, respectively) using SMRs induced by the motor imaginations [9].
Combining SMR-based BCIs and FES systems to help severely or completely paralyzed patients
is a new approach. This combination could not only restore motor functions [9], but it also has
the potential to facilitate neuroplasticity by performing MIs [10]. Furthermore, BCI-controlled FES
systems can provide the result of the imagined motion through proprioceptive sensory feedback [12].
For instance, [55] utilized the brain signals to trigger FES during index finger movement training, and
the participant showed the recovery of the volitional isolated index finger movements. In addition, [56]
applied an SMR-based BCI-controlled FES system to perform dorsiflexion of the paralyzed ankle in
a patient with stroke. The authors reported a promising result in that both the amplitude of EMG in
the affected tibialis anterior muscle and the range of movement at the ankle joint significantly increased
with the SMR-based BCI-controlled FES system in comparison with FES alone. [10] compared five
different interventions, such as (1) BCI-controlled FES, (2) EMG-controlled FES, (3) conventional push
button-controlled FES, (4) voluntary grasping, and (5) BCI-guided voluntary grasping with 10 healthy
participants to investigate the effects on neuroplasticity. More support for the use of SMRs induced
by MIs as promising interventions to improve motor function rehabilitation in stroke patients was
found in [57]. There were also many studies showing positive outcomes in using SMR-based BCIs for
the rehabilitation of patients with severe motor impairments [8,58].
computer control that is essential for synchronizing MI-evoked brain signals and FES-driven physical
feedback. Although some medical and research purpose-built FES systems support multi-channel
electrodes and real-time control, these systems are either expensive or not commercially available in
the United State.
3.2. Methods
3.2.1. Participants
A total of eight stroke patients (three females; mean age or M = 46.0 ± 13.5 years) were recruited
from local rehabilitation centers and clinics. All studies were reviewed and approved by the Institutional
Review Board of North Carolina State University. The participants received monetary compensation
for participation, and any participants who were displeased during the studies could cease at any time,
but no one had chosen to withdraw the study until the studies were complete. The inclusion criteria in
this study were (1) suffering from upper limb hemiparesis, weakness on one side of the body, the most
common impairment after stroke or TBI; (2) in the chronic state of stroke, the condition of a stroke
patient persisted without recurrent strokes for more than three months, to prevent the potential risk of
recurrent stroke, and (3) having normal sensitivity to feel pain and discomfort on the impaired forearm.
Then, participants were asked to imagine each motion depicted in the video clip as similar as possible.
The two different rhythmic motions used in visual guidance were intended to classify a 2-class MI
task in a single hand utilizing distinct SMR features evoked by different rhythmic MIs [65,66]. These
video clips were displayed on a 21-inch liquid-crystal display (LCD) monitor located approximately
42 inches in front of participants, and the centerline of the monitor was set within 10 degrees of eye
level to follow the guideline of viewing distance and screen size, and viewing angle for individuals,
respectively [67].
3.2.3. Procedure
FES electrodes were placed on the adequate coordinates with the user-specific FES parameters
identified [68]. Then, the anti-static wrist strap was applied to the opposite hand of the affected forearm
with the FES electrode to ground out the electrical stimulation. Afterward, the participants were then
asked to wear an EEG cap with 32 active EEG electrodes (g.tec medical engineering GmbH, Graz,
Austria). When the FES and BCI systems were ready, participants were given detailed explanations
of the synchronous experiment and were instructed on how to perform the MI tasks according to
the visual guideline shown on the monitor.
The experiment in Phase 1 consisted of five sessions with 24 trials per session, and there was
a three-minute break between sessions. Each trial lasted 12 s and consisted of a rest period (2.5 s),
a reference period with cues (2.5 s), an MI period (3 s), an FES initiation period (1 s), and a feedback
or FES period (3 s), as shown in Figure 1. During the rest period, the screen showed only the gray
background without any image, and an auditory cue (“take a rest”) was played at the beginning of
the rest period. Participants were required to remain calm during this period to stabilize EEG signals.
Then, the reference period was followed by a cross fixation screen and an auditory cue (“ready”) at
the beginning of the period, and the visual (text) cue and the auditory cue for either “slow grasping” or
“fast opening” were presented at the end of the reference period. The participants were asked to focus
on the screen to be ready for MI tasks. Afterward, the MI period was successively given to participants
with the visual guideline of SOG or FCO for three seconds. During this period, the participants were
asked to mimic MI tasks as close as possible by following the visual cues. After the MI period, the FES
initiation period lasted for 1 s, during which the FES system activated the same motion to the subjects’
affected forearm as presented in the visual guideline. In this period, the current amplitude of FES was
gradually increased by utilizing the ramp time function to improve user satisfaction. At the end of
the FES initiation period, the feedback period was successively given to the participants. In this period,
the participants were required to keep or stop the MI tasks given in the feedback period according to
auditory cues such as “keep imagination” or “stop imagination”, respectively.
In the MI period, participants conducted a total of 120 trials including 60 trials for each SOG and
FCO, while they performed 60 trials for either keep or stop MI tasks in the feedback period. All trials
were pseudo-randomized so that the number of trials for different conditions were balanced with each
combination. The total time of the experiment was around 40 min, and the completion time of Phase 1
was 60 min, including the setup time.
motion to the subjects’ affected forearm as presented in the visual guideline. In this period, the
current amplitude of FES was gradually increased by utilizing the ramp time function to improve
user satisfaction. At the end of the FES initiation period, the feedback period was successively given
to the participants. In this period, the participants were required to keep or stop the MI tasks given
inBrain
theSci.
feedback
2020, 10, period
512 according to auditory cues such as “keep imagination” or “stop imagination”,
7 of 27
respectively.
Figure1.
Figure Schematic illustration
1. Schematic illustrationofofthe experiment
the procedure.
experiment TextText
procedure. in theinblue
thebox indicates
blue the auditory
box indicates the
cue that played at the beginning of each period, and INI is an abbreviation of the Functional
auditory cue that played at the beginning of each period, and INI is an abbreviation of the Functional Electrical
Stimulation
Electrical (FES) initiation
Stimulation period. MI:
(FES) initiation Motor
period. Imagery.
MI: Motor Imagery.
T
XC1 XC1 T
XC2 XC2
CC1 = , CC2 = , Cc = CC1 + CC2 (1)
T
trace XC1 XC1 T
trace XC2 XC2
where trace (XCx XCx T ) is the sum of diagonal elements of (XCx XCx T ). Afterward, the composite spatial
√
covariance, C, and projection matrix, W, were calculated from Cc = CC1 + CC2 = Uc λUCT W = UT λUCT ,
where Uc is a matrix of normalized eigenvectors, and λ is eigenvalues. Then, the projection matrix W
transformed EEG signals into two CSP features.
Signal preprocessing: Once the brain signals were recorded, noise, artifacts, and other irrelevant
brain signals were filtered out. EEG signals were visually inspected for unexpected EEG contamination
usually caused by body movements and electrode drifts during the experiments, and those trials were
excluded from further processing. In addition, EEG signals were filtered by an automatic artifact
rejection toolbox from the EEGLAB. For details, see [70,71]. The rejection thresholds were set at
150 µV for the large fluctuation threshold and 3 standard deviations for the probability threshold
of entropy and kurtosis. The remaining trials were then band-pass filtered between 1 and 29 Hz to
remove irrelevant brain signals such as muscle artifacts. Afterward, principal component analysis
(PCA) was applied to the filtered EEG signals to project high-dimensional data (32 channels) into
lower dimensions while maintaining important brain features. After applying PCA, eigenvectors and
Brain Sci. 2020, 10, 512 8 of 27
eigenvalues of a covariance matrix were generated. Among all eigenvalues, the first few components
that could explain 95% of variance were selected with the assumption that the important brain features
are contained in those few components. This procedure not only addresses the high dimensionality
issue but also helps to speed up an independent component analysis (ICA) decomposition process by
reducing the dimension of the EEG signal.
The reduced dimensional EEG signals after the PCA procedure still mixed with irrelevant
independent components such as eye-blinking artifacts and visual stimulus-evoked potentials. Thus,
an ICA method was applied to retain only the relevant independent components by removing irrelevant
components. After the ICA decomposition process, the EEG signals were decomposed into maximally
independent components, and the artifact (irrelevant) components could be identified by plotting
each independent component in a component activity time course and a topographical map. After
removing the artifact-related components, EEG signals were projected back with the whitened data.
Finally, the surface Laplacian estimation was applied to the EEG signals between the reference and MI
periods (EEG subset 1) to enhance the spatial EEG traces of the ERD/ERS procedure [72].
Feature Extraction: After the signal preprocessing procedure, two feature extraction methods
were utilized, including ERD/ERS [73] and CSP methods [74]. The ERD/ERS method was applied
only to EEG subset 1, while the CSP was analyzed for all three EEG subsets. To compute ERD/ERS,
the preprocessed EEG subset including both SMR and reference periods was band-pass filtered with
a width of 4 Hz by FIR filter (For details, see [75]). Then, the EEG signals were divided into seven
sub-bands from the first sub-band bin, [1–5] Hz, to the seventh sub-band bin (for details, see [25–29]),
Hz, with 4 Hz intervals. Afterward, the EEG data of each sub-band bin were transformed from
time-domain data to frequency-domain data by using Fast Fourier Transform (FFT) with the Hamming
windowing method for 2.5-s data epoch (640 data points). Afterward, the relative band power of each
sub-band bin, ERD/ERS, was computed with the value of FFT in the reference period and that in the MI
period for each trial by following [76].
The three EEG subsets were analyzed with the sub-band CSP (SBCSP), which can project
a multidimensional EEG dataset into a low-dimensional subspace. SBCSP allows to decompose the EEG
subsets into seven sub frequency bands by following the same procedure utilized in the ERD/ERS
method. From each sub-band bin, the eigenvalues were calculated and utilized to determine the best
feature from the projected EEG matrix. Then, the features from SBCSP were utilized to increase
the subsequent classification accuracy by fusing the classification scores of each sub-band feature.
The advantages of SBCSP include (1) not requiring an exhaustive band selection procedure and (2)
allowing the use of multiple frequency bands at the same time [77].
Classification: Three different classification algorithms, Fisher’s LDA, SVM, and ensemble
methods, were tested to identify the best classification algorithm and classifier parameters. Many BCI
studies have reported good classification outcomes with LDA, as well as have validated these methods
as among the most suitable algorithms for SMR feature classification [78]. For the SVM classification
algorithm, the kernel function used quadratic components to cover non-linear characteristics in the brain
signal, and the least-squares method was applied to find the separating hyperplane. The ensemble
method by combining the outputs of the LDA and SVM classifiers was also evaluated to test hypothesis
1.3, and the final decision of the ensemble method was made with the weighted majority voting by
combining the decisions of LDA and SVM, as well as by combining the features of each classifier [79].
Each EEG subset was uniformly divided into 10 sub-datasets to apply a 10-fold cross-validation
method [80]. Nine sub-datasets were used as training datasets to build a classification weight vector,
and the remaining one sub-dataset was utilized as a test dataset to calculate the classification accuracy.
was not being applied; (2) ACT period: classifying a 2-class MI task in a single hand, and (3) FES
period: decoding the existence of SMRs to control FES when FES was being applied. The classification
type had three levels for all periods such as LDA, SVM, and ensemble methods.
Accuracy (%) was measured as a dependent variable (DV), which is the ratio of the correct
classifications over the number of trials [45]. The results from 10-fold cross-validation for each period
were averaged to calculate the accuracy with three classification types [80]. Since the experiment
in Phase 1 was a balanced 3 × 3 within-subjects factorial design with two IVs (period type and
classification type) with participant as a blocking variable and one DV (accuracy), a univariate
ANOVA was used. Prior to conducting two-way ANOVAs on the accuracy data, the assumptions
of both homoscedasticity and normality of residuals were assessed in JMP® (SAS institute Inc.,
Cary, NC, USA). The results of the Shapiro–Wilk test for normality (W = 0. 982593, p = 0.4207) and
Brown–Forsythe test for homoscedasticity (F8, 54 = 0.7150, p = 0.6773) showed that the data did not
violate the ANOVA assumptions.
The results showed that the average accuracies between three periods were significantly different
(F2, 56 = 47.8111; p < 0.00001; ηp2 = 0.63066), but the main effect of classifier types and interaction were
not significant. Tukey’s honest significance test (HSD) was tested to determine significant differences in
the main effects of the MI period. The results showed that the average accuracy of the SMR period was
significantly higher than that of the ACT period, and that of the ACT period was significantly higher
than that of the FES period. Furthermore, the results of the t-test showed that the average accuracy
of the ACT period (68.82%; t23 = 4.10, p < 0.001) was significantly higher than the true chance level
(57.43%), but the FES period (59.16%; t23 = 1.17, p = 0.2558) was not significant. Tukey’s HSD was also
tested to determine significant differences in the main effects of classifier type. Although the average
accuracy of the ensemble method was higher than others, the results of the statistical analysis showed
that the average accuracies between classifiers were not significantly different.
3.4. Discussion
As expected, the results of statistical analysis showed that the classification accuracy of a 2-class
MI task was significantly higher than the true chance level. Therefore, the hypothesis H1.1 was rejected
in favor of the alternative hypothesis, which states that the classification accuracy to classify a 2-class
Brain Sci. 2020, 10, 512 10 of 27
MI task in a single hand is significantly higher than the true chance level. This result could be extended
to any SMR-based BCI system, as well as other SMR-based FES-BCI systems. The results of statistical
analysis also showed that the accuracy of the FES period was not significantly higher than the true
chance level. Therefore, the hypothesis H1.2 was not rejected. The reasons for these results are as
follows. First, the duration of the FES period might be too short to stabilize the excited brain signal due
to electrical stimulation. Second, electrical artifacts from the FES might be not completely grounded by
the anti-static wrist strap applied to the opposite hand of the affected forearm with the FES electrode.
The results of the classification type showed that the average accuracy of the ensemble method
was higher than other methods, but statistical analysis showed that the difference between classification
types was not significant. One possible reason is that the accuracy of classification among subjects
showed a large variation due to the difference in individual ability. To address this issue, the accuracy
of classification was standardized within participants, and ANOVA analysis was conducted again.
The results of ANOVA showed that the standardized accuracy between classification types was
significantly different (F2, 56 = 3.347; p = 0.0424; ηp2 = 0.10677). Therefore, the modified hypothesis of
H1.3 was rejected, which means that the standardized classification accuracy of the ensemble method
is significantly higher than that of LDA. Finally, two classification features including the user-specific
classifier parameter set and a probability weight for the ensemble method were constructed.
4.2. Methods
4.2.1. Participants
The same patients who participated in Phase 1 continued this experiment either after a 20-min
break or when the participants felt ready. Any patients who were displeased with the experiment
could cease at any time, but no one had chosen to withdraw.
group are extracted. Then, classifier parameters are generated to best distinguish each group based on
the features. The second mode is an asynchronous, self-paced experiment where the users perform
a task at their own pace rather than following the cues. The advantage of the asynchronous experiment
is that the BCI system can be more realistic and flexible, because it allows the user to perform the desired
operation at his/her own pace [82]. It would be very useful for patients to be able to control the impaired
body solely through MIs. However, since it is difficult to know when the user performed MI tasks,
the asynchronous mode has a critical issue such that the classification accuracy is low due to the high
false positive ratio [83]. Due to the low accuracy of the MI tasks because of the high false positive ratio,
it is somewhat difficult to perform experiments on patients in a completely asynchronous situation.
As a result, asynchronous BCI research has not been studied as much compared to synchronous BCI
research [60].
More training in synchronous experiments could help to solve this issue, but long-lasting BCI
training is not easy for participants with impaired physical conditions (and/or mental conditions such as
attention deficits after stroke and TBI), since it requires participants to sit down and focus on repetitive
mental tasks while minimizing movement. Furthermore, the results of synchronous experiments are
not always guaranteed to be the same results as those of asynchronous experiments. This is due to
cue differences (Yes versus No), analysis time differences (fixed versus variable), and the nature of
brain signals that have non-stationarity and inter- and intra-variability [84]. Alternatively, we can
consider increasing the accuracy through more practice in the same environment as the asynchronous
mode, but there are also the following issues. First, it is difficult to know when the patient has
successfully performed MIs, as knowing this information is essential to categorize the EEG signals into
the two groups (Yes-SMR versus No-SMR) to initiate the FES system. Second, if the errors, either false
positive or false negative, are accumulated, the participants will be easily confused and frustrated.
Finally, an additional time required to correct the erroneous movements will increase the duration
of the experiment, and we should take into consideration that prolonged FES may increase muscle
fatigue in patients.
Thus, Phase 2 used a semi-asynchronous mode for performing a fixed sequence. This mode has
three important features that combine the advantages of both synchronous and asynchronous modes
as follows: (1) a fixed sequence of tasks to clarify what the user needed to do, (2) error-free results to
prevent participants from being confused or exhausted, and (3) initiating the FES system by detecting
SMRs similar to the asynchronous mode. Under the proposed semi-asynchronous mode, participants
would be free from confusion and frustration while using the SMR-based BCI-controlled FES system at
their own pace. Furthermore, the experimenter can enhance the classification performance by adjusting
the classifier parameters using the most recently measured brain signals under this mode, which is
also known as adaptive machine learning [85,86].
4.2.3. Procedure
Participants were asked to conduct the fixed sequence MI tasks according to a specific goal.
The BCI tasks consisted of (1) opening the hand with FCO MIs similar to the previous study; (2)
stopping the FES system by not imagining any action similar to “stop imagination”; (3) grasping a ball
with SOG MIs; (4) moving the ball to the target position without going through the BCI system, and (5)
holding the ball on the target position by continuously imagining SOG. In this experiment, patients
were asked to perform all tasks as quickly as possible, except the ‘holding the ball’ task, which should
last as long as possible, using the appropriate MIs trained in the previous experiment (see Figure 2).
In Figure 2, the colored boxes represent the tasks that the participant should perform through
the BCI system, and the white boxes represent the tasks that the participant was asked to perform
without using the BCI system. The green boxes on the left indicate MI tasks without FES, and the blue
boxes in the middle show the tasks performed under FES. The following describes each step that
participants were asked to perform in detail.
last as long as possible, using the appropriate MIs trained in the previous experiment (see Figure 2).
In Figure 2, the colored boxes represent the tasks that the participant should perform through
the BCI system, and the white boxes represent the tasks that the participant was asked to perform
without using the BCI system. The green boxes on the left indicate MI tasks without FES, and the blue
Brain
boxes Sci.in the10,middle
2020, 512 show the tasks performed under FES. The following describes each step12 of 27
that
participants were asked to perform in detail.
Figure 2.
Figure 2. Sequence
Sequence of
of the
the predefined
predefined tasks
tasks in
in Phase
Phase 2.
2. BCI:
BCI: Brain-Computer
Brain-Computer Interface.
Interface.
Step
Step 1—Open
1—Opena hand:a hand:First, participants
First, werewere
participants askedasked
to perform the opening
to perform MI task byMI
the opening imagining
task by
imagining
FCO trainedFCO trained
in the in the
previous previous experiment.
experiment. Once the BCIOnce systemthedetected
BCI system detected anSMR,
an FCO-related FCO-related
the FES
SMR, the
system wasFES systemfor
activated wasanactivated for an opening
opening motion of 5 s. If motion
the systemof 5did
s. If
notthe system
detect the did not detectSMR
FCO-related the
FCO-related
in the maximum SMRanalysis
in the maximum analysis
time (6 s), then time (6 was
the system s), then the systemstarted
automatically was automatically started so
so that the experiment
that the
time didexperiment
not increase time did not increase
excessively. excessively.
Afterward, Afterward,were
the participants the participants
asked to movewere asked
their to move
hand near
atheir
ball hand
to be near
readya to
ball to be ready to grab.
grab.
Step 2—Stop FES: After five seconds, the participants were asked to stop the FES system as soon
as possible without imagining any motion similar to “stop imagination” in Phase 2. When the BCI
system detected the absence of MI, the FES system stopped working. To prevent excessive muscle
fatigue and discomfort, if participants could not stop the FES system in the maximum analysis time,
the system automatically stopped operating. Then, participants were given a 10-s break to stabilize
excited brain activity from the physical movements and electrical stimulation, and the BCI system also
stopped classifying brain signals during the break.
Step 3—Grab a ball: After the break, the participants were asked to conduct the SOG MIs to grab
the ball. If an SOG-related SMR was detected, the FES system was activated for grasping motion for 5 s,
and they were asked to move the ball to the target position within 5 s. Similar to Step 1, if the system
did not detect the SOG-related SMR in the maximum analysis time, then the system automatically
initiated a grasping motion.
Step 4—Hold the ball: After the participants had moved the ball over to the target position,
they tried to keep holding the ball for another 5 s. To hold the ball, the grasping SMR should be
maintained when FES was activated for grasping. If the grasping SMR was successfully maintained
during the maximum analysis time, the FES system automatically stopped for safety. This was the task
sequence of this experiment necessary to complete one trial.
Participants were asked to conduct two sets of the sequential tasks in each trial, which lasted until
all tasks were finished. A total of 20 trials were conducted during the experiment. The completion
time varied due to the different performance levels between participants, but the average completion
time was around 40 min.
Study 1, including the SMR period (to detect the SMR features to initiate the FES system), the ACT
period (classifying different MIs including grasping and opening), and the FES period (to either keep
or stop the FES system). One difference was that the incoming signal was the same, but the role
was changed according to the current goal. For example, if the current status is to try to perform
opening a hand, then the current period is defined as the SMR period. If the SMR feature is detected,
then the role of the same EEG dataset is changed to the ACT period. Therefore, detecting SMRs and
classifying a 2-class MI task are performed sequentially. However, if SMR-related brain features are
not detected, the current role is not changed, and the SMR period is maintained. On the contrary, if
a target role of either grasping or opening is not detected during the ACT period, then the current
status is set back to the SMR period. When the BCI system detects only the target role from the ACT
period, the FES system is activated according to the decision from the BCI system.
Adaptive Learning: When participants completed the first 10 trials, the adaptive learning method
was applied to adjust classifier parameters. Many different adaptive learning methods are available,
and the advantages and computational time of each method vary. More elaborate adaptive learning
methods might yield higher accuracy, but they require more computational time. Adaptive learning
methods can be divided into two groups, supervised and unsupervised learning. As [85] recommended
due to ease and good performance, the Pooled mean (PMean) adaptive estimation method was
chosen. Since this method is the unsupervised adaptive learning method, it can also be utilized in
the asynchronous mode. The PMean estimation can be determined as follows:
where W0t = the bias of the classifier at trial, η = the updating coefficient (0.1), and xt = the new feature
obtained at trial t.
period, it is a delay rather than an error requiring additional work to be corrected. In addition, due to
the nature of the asynchronous experiment, it is difficult to distinguish whether the delay is intended
due to a failure to implement MIs. However, from another point of view, it is difficult to define this as
a successful task performance if the delay caused by MI induction failure is long enough to affect BCI
performance. Therefore, the accuracy in this experiment was defined as follows. (1) If an incorrect
decision was made during the SMR period (i.e., no SMR was detecting), then it belongs to a delay until
the maximum analysis time (6 s). (2) If an incorrect decision was made during ACT (i.e., the target
action was ‘opening’, but classified as ‘grasping’, and vice versa), then it belongs to an error. (3) If an
incorrect decision was made during the FES period (i.e., the action should be ‘keep’, but it was classified
as ‘stop’, and vice versa), then it belongs to an error. The structure of the classification algorithms
utilized in this phase also agree on this argument, because even if it fails to detect the SMR function,
it is not necessary to correct this error, but only a delay of 1s occurs. (3) Information Transfer Rate:
Many BCI studies have utilized Information Transfer Rate (ITR) to evaluate the performance of BCI
systems, which can convey the amount of information per time (bits/minute) in terms of the accuracy
and speed [90].
Subjective Assessment: Participants were asked to complete the NASA Task Load Index (NASA-TLX)
questionnaire to measure the subjective mental workload, which is the same as in Phase 2 [91]. Then,
the results of studies were analyzed to compare changes in the subjective mental workload between
the two experiments.
4.5. Results
Table 3 shows the average completion rate for each IV. The results showed that the average
completion rate of the No-FES period was higher than that of the Yes-FES period. Furthermore,
the average completion rate after applying the adaptive learning method was higher than before.
The accuracy of each task is shown in Table 4. The results showed that the accuracy of the No-FES
period was a little higher than that of the Yes-FES period, but the differences were not large. There was
some evidence that the classifiers might be biased to a certain direction because Subject S05 tended to
say ‘keep FES’, while Subject S04 tended to say ‘stop FES’.
Table 5 shows the accuracy of each IV, and the results showed that the accuracy of each FES
condition increased after applying learning. It indicates that the adaptive learning procedure enhanced
the identifiability of both classification algorithms. Between two periods, the improvement of
the No-FES period was higher than that of the Yes-FES period.
Brain Sci. 2020, 10, 512 16 of 27
Table 6 shows the ITR of each IV, and the ITR values increased a little bit after applying the adaptive
learning method, but the difference was great. The results of ITR also indicated that the performance
of the proposed SMR-based BCI system fell into the average ITR ranges (10–25 bit/min) [7].
An ANOVA was also performed to test the main effects of IVs, the existence of FES, and the learning
type and interactions between IVs with participant as a blocking variable on the completion rate and
ITR, separately. An ANOVA was also performed to test the main effects of IVs, the existence of FES, and
the learning type, and interactions between IVs and the participants as a blocking variable on accuracy.
There were significant differences in learning type (F1, 21 = 8.2273; p = 0.0092; ηp2 = 0.28149), but
the main effect of the FES existence type and interaction effect were not significant. In addition, Tukey’s
HSD test showed that the average accuracy during the No-FES period after applying adaptive learning
was significantly higher than that during the Yes-FES period before applying adaptive learning.
4.5.2. Workload
To analyze the subjective mental workload changes between the two BCI experiments in Phase 1
and 2, the participants were asked to answer the NASA-TLX questionnaire after each experiment [91].
To test the significance of the differences between the experiments, a two-tailed Student’s t-test was
performed with the differences (scores of the second experiment—score of the first experiment) of each
question. Table 7 shows that mental demand and frustration increased significantly, while performance
significantly decreased compared to the first experiment.
Brain Sci. 2020, 10, 512 17 of 27
4.6. Discussion
5. General Discussion
seconds, and the y-axis displays the current FES status. Black boxes indicate the waiting time given
at
thethe beginning
beginning oror between
between sequences.Green
sequences. Greenboxes
boxesindicate
indicatethe
thetime
timewhen
when the
the BCI
BCI system
system does
does not
not
classify
classify in the 5 s given immediately after performing the MI task between periods. Red boxes mean
in the 5 s given immediately after performing the MI task between periods. Red boxes mean
incorrect classificationresults,
incorrect classification results,and
andblue
blueboxes
boxesimply
implythethe time
time that
that successfully
successfully performed
performed the the keep-
keep-FES
FES task. This box should appear for 5 s, meaning it was done perfectly without mistakes
task. This box should appear for 5 s, meaning it was done perfectly without mistakes (5 correct (5 correct
decisions),
decisions), unlike
unlike the
the red
red box.
box.
(a) (b)
the adaptive learning method under semi-asynchronous mode. Then, the proposed experiment mode
plays an important role to bridge the gap between the two modes. Furthermore, this approach can not
only improve task performance by minimizing the misclassification rate but can also enhance user
performance by increasing self-esteem and reducing frustration.
accuracies during the FES period for both asynchronous and semi-asynchronous experiments were
not higher than the true chance level, and it implies that the proposed SMR-based BCI system could
not distinguish between voluntary MI-evoked SMRs and FES-driven passive-movement-evoked
SMRs. There are three possible reasons to explain this low accuracy. First, when FES was activated,
the electrical artifacts might be not completely grounded by the anti-static wrist strap applied to
the opposite hand of the affected forearm. Thus, it is possible that electrical stimulation was mixed
with the EEG signals, making it difficult to classify. Second, these passive movements driven by
FES might have prevented the patients from stopping the MI tasks, as some participants mentioned
that it was difficult to stay calm without movement-related imagination when FES was activated.
Finally, although the participants were asked not to perform any voluntary physical movements
during the experiments, they might simultaneously perform spontaneous movements by following
the FES-driven motion. Since EMG signals were not measured in this study, it was not known whether
the motion had progressed. Therefore, this limitation should be solved through further study.
Although the classification accuracy of the No-FES period after applying adaptive learning
under semi-asynchronous mode was significantly higher than the true chance level, it did not achieve
the threshold level (70%) for an effective control through a BCI system [92]. Moreover, the adaptive
learning method significantly increased the accuracy similar to other studies [61], but these results might
be due to an order effect that the participants were accustomed to the semi-asynchronous experiment
while continuing to conduct the experiment. The other limitation was that adequate FES electrode
placement was altered during supination and pronation, leading to sharp muscle pain. Participants
were asked to minimize supination and pronation to prevent any muscle pain and discomfort; however,
this restriction made their movement unnatural. As [104] reviewed, the array types of FES electrode are
available, which could help to solve this issue by synchronizing between supination (pronation) and
the electrode onset in real time. Finally, this study did not consider the reinforcement or attenuation of
SMRs over time. It might be possible to identify the change of SMRs as the time progressed while
conducting a long-time experiment, and the relationship between the time and MI task performance
could be evaluated to determine the proper duration of the SMR-based BCI experiment with respect
to the fatigue and tiredness of participants. In this way, task performance can be maximized, and
attention deterioration can be minimized.
Author Contributions: Conceptualization, C.S.N. and I.C.; methodology, C.S.N. and I.C.; validation, I.C. and
C.S.N.; formal analysis, I.C. and C.S.N.; investigation, I.C., G.H.K. and S.L.; data curation, I.C.; writing—original
draft preparation, C.S.N. and I.C.; writing—review and editing, C.S.N., I.C., G.H.K. and S.L.; visualization, I.C.;
supervision, C.S.N.; project administration, C.S.N.; funding acquisition, C.S.N. All authors have read and agreed
to the published version of the manuscript.
Funding: This work was supported in part by the National Science Foundation (NSF) under Grant Number
IIS-1421948. Any opinions, findings, and conclusions or recommendations expressed in this material are those of
the authors and do not necessarily reflect the views of the NSF.
Acknowledgments: We would like to thank all those who have helped in carrying out the research.
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Van Delden, A.E.Q.; Peper, C.E.; Kwakkel, G.; Beek, P.J. A systematic review of bilateral upper limb training
devices for poststroke rehabilitation. Stroke Res. Treat. 2012, 2012, 1–17. [CrossRef] [PubMed]
2. Iosa, M.; Hesse, S.; Oliviero, A.; Paolucci, S. New technologies for stroke rehabilitation. Stroke Res. Stroke Res.
Treat. 2013, 2013, 2–3. [CrossRef] [PubMed]
3. Peckham, P.H.; Knutson, J.S. Functional electrical stimulation for neuromuscular applications. Annu. Rev.
Biomed. Eng. 2005, 7, 327–360. [CrossRef] [PubMed]
4. Sujith, O.K. Functional electrical stimulation in neurological disorders. Eur. J. Neurol. 2008, 15, 437–444.
[CrossRef]
5. Johnson, L.A.; Fuglevand, A.J. Mimicking muscle activity with electrical stimulation. J. Neural Eng. 2011, 8,
16009. [CrossRef]
Brain Sci. 2020, 10, 512 23 of 27
6. Popovic, M.R.; Curt, A.; Keller, T.; Dietz, V. Functional electrical stimulation for grasping and walking:
Indications and limitations. Spinal Cord 2001, 39, 403–412. [CrossRef]
7. Wolpaw, J.R.; Birbaumer, N.; McFarland, D.J.; Pfurtscheller, G.; Vaughan, T.M. Brain-computer interfaces for
communication and control. Clin. Neurophysiol. 2002, 113, 767–791. [CrossRef]
8. Daly, J.J.; Wolpaw, J.R. Brain-computer interfaces in neurological rehabilitation. Lancet Neurol. 2008, 7,
1032–1043. [CrossRef]
9. Pfurtscheller, G.; Müller, G.R.; Pfurtscheller, J.; Gerner, H.J.; Rupp, R. “Thought”—Control of functional
electrical stimulation to restore hand grasp in a patient with tetraplegia. Neurosci. Lett. 2003, 351, 33–36.
[CrossRef]
10. McGie, S.C.; Zariffa, J.J.; Popovic, M.R.; Nagai, M.K. Short-term neuroplastic effects of brain-controlled and
muscle-controlled electrical stimulation. Neuromodulation 2015, 18, 233–240. [CrossRef]
11. Rohm, M.; Muller-Putz, G.R.; Kreilinger, A.; Von Ascheberg, A.; Rupp, R. A hybrid-Brain Computer Interface
for control of a reaching and grasping neuroprosthesis. Biomed. Tech. 2010, 55 (Suppl. S1), 1–4.
12. Hara, Y. Neurorehabilitation with new functional electrical stimulation for hemiparetic upper extremity in
stroke patients. J. Nippon. Med Sch. 2008, 75, 4–14. [CrossRef] [PubMed]
13. Mozaffarian, D.; Benjamin, E.J.; Go, A.S.; Arnett, D.K.; Blaha, M.J.; Cushman, M.; Das, S.R.; De Ferranti, S.;
Despres, J.-P.; Fullerton, H.J.; et al. Heart Disease and Stroke Statistics—2016 Update: A Report From
the American Heart Association. Circulation 2015, 133. [CrossRef] [PubMed]
14. Hirtz, D.; Thurman, D.; Gwinn-Hardy, K.; Mohamed, M.; Chaudhuri, A.; Zalutsky, R. How common are
the “common” neurologic disorders? Neurology 2007, 68, 326–337. [CrossRef] [PubMed]
15. Pereira, S.; Teasell, R.; Graham, R.; Salter, K. Rehabilitation of Severe Stroke. Evid. Based Rev. Stroke Rehabil.
Module 2013, 23, 1–32.
16. Miller, E.L.; Murray, L.; Richards, L.; Zorowitz, R.D.; Bakas, T.; Clark, P.; Billinger, S.A. Comprehensive
overview of nursing and interdisciplinary rehabilitation care of the stroke patient: A scientific statement
from the American heart association. Stroke 2010, 41, 2402–2448. [CrossRef]
17. Wissel, J.; Schelosky, L.D.; Scott, J.; Christe, W.; Faiss, J.; Mueller, J. Early development of spasticity following
stroke: A prospective, observational trial. J. Neurol. 2010, 257, 1067–1072. [CrossRef]
18. Langhorne, P.; Coupar, F.; Pollock, A. Motor recovery after stroke: A systematic review. Lancet Neurol. 2009,
8, 741–754. [CrossRef]
19. Duncan, P.W.; Zorowitz, R.D.; Bates, B.; Choi, J.Y.; Glasberg, J.J.; Graham, G.D.; Katz, R.C.; Lamberty, K.;
Reker, D. Management of Adult Stroke Rehabilitation Care. Stroke 2005, 36, e100–e143. [CrossRef]
20. Matsumoto, S.; Shimodozono, M.; Etoh, S.; Shimozono, Y.; Tanaka, N.; Kawahira, K. Beneficial effects of
footbaths in controlling spasticity after stroke. Int. J. Biometeorol. 2010, 54, 465–473. [CrossRef]
21. Kawashima, N.; Popovic, M.R.; Zivanovic, V. Effect of Intensive Functional Electrical Stimulation Therapy
on Upper-Limb Motor Recovery after Stroke: Case Study of a Patient with Chronic Stroke. Physiother. Can.
2013, 65, 20–28. [CrossRef] [PubMed]
22. Quandt, F.; Hummel, F.C. The influence of functional electrical stimulation on hand motor recovery in stroke
patients: A review. Exp. Transl. Stroke Med. 2014, 6, 9. [CrossRef] [PubMed]
23. Lynch, C.L.; Popovic, M.R. Functional electrical stimulation. IEEE Control Syst. 2008, 28, 40–50.
24. Levin, M.F.; Hui-Chan, C.W. Relief of hemiparetic spasticity by TENS is associated with improvement in
reflex and voluntary motor functions. Electroencephalogr. Clin. Neurophysiol. Potentials Sect. 1992, 85, 131–142.
[CrossRef]
25. Sabut, S.K.; Sikdar, C.; Kumar, R.; Mahadevappa, M. Functional electrical stimulation of dorsiflexor
muscle: Effects on dorsiflexor strength, plantarflexor spasticity, and motor recovery in stroke patients.
Neurorehabilitation 2011, 29, 393–400. [CrossRef]
26. Zhang, D.; Guan, T.H.; Widjaja, F.; Ang, W.T. Functional electrical stimulation in rehabilitation engineering: A
Survey. In i-CREATe ’07: Proceedings of the 1st International Convention on Rehabilitation Engineering & Assistive
Technology in Conjunction with 1st Tan Tock Seng Hospital Neurorehabilitation Meeting; ACM Press: New York,
NY, USA, 2007; pp. 221–226.
27. Kottink, A.I.R.; Oostendorp, L.J.; Buurke, J.H.; Nene, A.V.; Hermens, H.J.; Ijzerman, M.J. The Orthotic Effect
of Functional Electrical Stimulation on the Improvement of Walking in Stroke Patients with a Dropped Foot:
A Systematic Review. Artif. Organs 2004, 28, 577–586. [CrossRef]
Brain Sci. 2020, 10, 512 24 of 27
28. Jackson, A.; Mavoori, J.; Fetz, E.E. Long-term motor cortex plasticity induced by an electronic neural implant.
Nature 2006, 444, 56–60. [CrossRef]
29. Teasell, R.; Bayona, N.A.; Bitensky, J. Plasticity and Reorganization of the Brain Post Stroke. Top. Stroke
Rehabil. 2005, 12, 11–26. [CrossRef]
30. Papachristos, A. Functional Electrical Stimulation in Paraplegia. In Topics in Paraplegia; InTech: Rijeka,
Croatia, 2014.
31. Young, B.M.; Nigogosyan, Z.; Nair, V.A.; Walton, L.M.; Song, J.; Tyler, M.E.; Edwards, D.F.; Caldera, K.;
Sattin, J.A.; Williams, J.C.; et al. Case report: Post-stroke interventional BCI rehabilitation in an individual
with preexisting sensorineural disability. Front. Neuroeng. 2014, 7, 18. [CrossRef]
32. Chae, J.; Sheffler, L.R.; Knutson, J.S. Neuromuscular Electrical Stimulation for Motor Restoration in
Hemiplegia. Top. Stroke Rehabil. 2008, 15, 412–426. [CrossRef]
33. Lawrence, M. Transcutaneous Electrode Technology for Neuroprostheses. Ph.D. Thesis, Federal Technical
University, Zurich, Switzerland, 2009.
34. Mangold, S.; Keller, T.; Curt, A.; Dietz, V. Transcutaneous functional electrical stimulation for grasping in
subjects with cervical spinal cord injury. Spinal Cord 2004, 43, 1–13. [CrossRef] [PubMed]
35. Draganski, B.; Gaser, C.; Busch, V.; Schuierer, G.; Bogdahn, U.; May, A. Neuroplasticity: Changes in grey
matter induced by training. Nature 2004, 427, 311–312. [CrossRef] [PubMed]
36. Lourenção, M.I.P.; Battistella, L.R.; De Brito, C.M.M.; Tsukimoto, G.R.; Miyazaki, M.H. Effect of biofeedback
accompanying occupational therapy and functional electrical stimulation in hemiplegic patients. Int. J.
Rehabil. Res. 2008, 31, 33–41. [CrossRef] [PubMed]
37. Brouwer, A.-M.; Van Erp, J.B.F. A tactile P300 brain-computer interface. Front. Mol. Neurosci. 2010, 4, 19.
[CrossRef] [PubMed]
38. Mak, J.N.; Wolpaw, J.R. Clinical Applications of Brain-Computer Interfaces: Current State and Future
Prospects. IEEE Rev. Biomed. Eng. 2009, 2, 187–199. [CrossRef] [PubMed]
39. Käthner, I.; Ruf, C.A.; Pasqualotto, E.; Braun, C.; Birbaumer, N.; Halder, S. A portable auditory P300
brain–computer interface with directional cues. Clin. Neurophysiol. 2013, 124, 327–338. [CrossRef]
40. Barsi, G.I.; Popović, D.B.; Tarkka, I.M.; Sinkjær, T.; Grey, M.J. Cortical excitability changes following grasping
exercise augmented with electrical stimulation. Exp. Brain Res. 2008, 191, 57–66. [CrossRef]
41. Pfurtscheller, G.; Solis-Escalante, T.; Ortner, R.; Linortner, P.; Müller-Putz, G.R. Self-Paced Operation of an
SSVEP-Based Orthosis with and Without an Imagery-Based “Brain Switch”: A Feasibility Study Towards
a Hybrid BCI. IEEE Trans. Neural Syst. Rehabil. Eng. 2010, 18, 409–414. [CrossRef]
42. Holz, E.M.; Botrel, L.; Kübler, A. Independent home use of Brain Painting improves quality of life of two
artists in the locked-in state diagnosed with amyotrophic lateral sclerosis. Brain Comput. Interfaces 2015, 2,
117–134. [CrossRef]
43. Li, Y.; Nam, C.S. A Collaborative Brain-Computer Interface (BCI) for ALS Patients. Proc. Hum. Factors Ergon.
Soc. Annu. Meet. 2015, 59, 716–720. [CrossRef]
44. Nam, C.S.; Lee, J.; Bahn, S.; Li, Y.; Choi, I. Brain-Computer Interface Supported Collaborative Work. In
Proceedings of the 5th International Brain-Computer Interface Meeting, Abingdon, UK, 3–7 June 2014.
45. Nam, C.S.; Moore, M.; Choi, I.; Li, Y. Designing Better, Cost-Effective Brain–Computer Interfaces. Ergon. Des.
Q. Hum. Factors Appl. 2015, 23, 13–19. [CrossRef]
46. Bhattacharyya, S.; Khasnobish, A.; Ghosh, P.; Mazumder, A.; Tibarewala, D.N. A Review on Brain Imaging
Techniques for BCI Applications. In Medical Imaging: Concepts, Methodologies, Tools and Applications; IGI
Global: Hershey, PA, USA, 2015.
47. Homer, M.L.; Nurmikko, A.V.; Donoghue, J.P.; Hochberg, L.R. Implants and Decoding for Intracortical Brain
Computer Interfaces. Annu. Rev. Biomed. Eng. 2013, 15, 383–405. [CrossRef] [PubMed]
48. Amiri, S.; Rabbi, A.; Azinfar, L.; Fazel-Rezai, R.; Asadpour, V. A Review of P300, SSVEP, and Hybrid P300 /
SSVEP Brain- Computer Interface Systems. Brain Comput. Interfaces Syst. 2013, 2013, 1–8.
49. Beisteiner, R.; Höllinger, P.; Lindinger, G.; Lang, W.; Berthoz, A. Mental representations of movements. Brain
potentials associated with imagination of hand movements. Electroencephalogr. Clin. Neurophysiol. Potentials
Sect. 1995, 96, 183–193. [CrossRef]
50. Müller, G.R.; Neuper, C.; Rupp, R.; Keinrath, C.; Gerner, H.J.; Pfurtscheller, G. Event-related beta EEG
changes during wrist movements induced by functional electrical stimulation of forearm muscles in man.
Neurosci. Lett. 2003, 340, 143–147. [CrossRef]
Brain Sci. 2020, 10, 512 25 of 27
51. Neuper, C.; Wörtz, M.; Pfurtscheller, G. ERD/ERS patterns reflecting sensorimotor activation and deactivation.
Prog. Brain Res. 2006, 159, 211–222.
52. Pfurtscheller, G. Graphical display and statistical evaluation of event-related desynchronization (ERD).
Electroencephalogr. Clin. Neurophysiol. 1977, 43, 757–760. [CrossRef]
53. Pfurtscheller, G. Event-related synchronization (ERS): An electrophysiological correlate of cortical areas at
rest. Electroencephalogr. Clin. Neurophysiol. 1992, 83, 62–69. [CrossRef]
54. Zhang, Y.; Nam, C.S.; Zhou, G.; Jin, J.; Wang, X.; Cichocki, A. Temporally constrained sparse group spatial
patterns for motor imagery BCI. IEEE trans. cyber. 2018, 49, 3322–3332. [CrossRef]
55. Daly, J.J.; Cheng, R.; Rogers, J.; Litinas, K.; Hrovat, K.; Dohring, M. Feasibility of a New Application of
Noninvasive Brain Computer Interface (BCI): A Case Study of Training for Recovery of Volitional Motor
Control After Stroke. J. Neurol. Phys. Ther. 2009, 33, 203–211. [CrossRef]
56. Takahashi, M.; Takeda, K.; Otaka, Y.; Osu, R.; Hanakawa, T.; Gouko, M.; Ito, K. Event related
desynchronization-modulated functional electrical stimulation system for stroke rehabilitation: A feasibility
study. J. Neuroeng. Rehabil. 2012, 9, 56. [CrossRef] [PubMed]
57. Sharma, N.; Pomeroy, V.M.; Baron, J.C. Motor imagery: A backdoor to the motor system after stroke? Stroke
2006, 37, 1941–1952. [CrossRef] [PubMed]
58. De Vries, S.; Mulder, T. Motor imagery and stroke rehabilitation: A critical discussion. J. Rehabil. Med. 2007,
39, 5–13. [CrossRef] [PubMed]
59. Krusienski, D.J.; Grosse-Wentrup, M.; Galán, F.; Coyle, D.; Miller, K.J.; Forney, E.; Anderson, C.W. Critical
issues in state-of-the-art brain–computer interface signal processing. J. Neural Eng. 2011, 8, 25002. [CrossRef]
[PubMed]
60. Lotte, F.; Congedo, M.; Lecuyer, A.; Lamarche, F.; Arnaldi, B. A review of classification algorithms for
EEG-based brain–computer interfaces. J. Neural Eng. 2007, 4, R1–R13. [CrossRef]
61. Vidaurre, C.; Blankertz, B. Towards a Cure for BCI Illiteracy. Brain Topogr. 2009, 23, 194–198. [CrossRef]
62. Guger, C.; Edlinger, G.; Harkam, W.; Niedermayer, I.; Pfurtscheller, G. How many people are able to operate
an eeg-based brain-computer interface (BCI)? IEEE Trans. Neural Syst. Rehabil. Eng. 2003, 11, 145–147.
[CrossRef]
63. Gatti, R.; Tettamanti, A.; Gough, P.M.; Riboldi, E.; Marinoni, L.; Buccino, G. Action observation versus motor
imagery in learning a complex motor task: A short review of literature and a kinematics study. Neurosci. Lett.
2013, 540, 37–42. [CrossRef]
64. Schuster, C.; Hilfiker, R.; Amft, O.; Scheidhauer, A.; Andrews, B.; Butler, J.A.; Kischka, U.; Ettlin, T. Best
practice for motor imagery: A systematic literature review on motor imagery training elements in five
different disciplines. BMC Med. 2011, 9, 75. [CrossRef]
65. Choi, I.; Bond, K.; Krusienski, D.; Nam, C.S. Effects of Off-Site Attention on SSSEP Amplitude. In Proceedings
of the 6th International Brain-Computer Interface Meeting, Pacific Grove, CA, USA, 30 May–3 June 2016.
66. Gu, Y.; Dremstrup, K.; Farina, D. Single-trial discrimination of type and speed of wrist movements from EEG
recordings. Clin. Neurophysiol. 2009, 120, 1596–1600. [CrossRef]
67. Ahlstrom, V.; Kudrick, B. Human Factors Criteria for Displays: A human Factors Design Standard—Update of
Chapter 5; FAA William J. Hughes Technical Centre: Atlantic City, NJ, USA, 2007.
68. Popović, D.B. Advances in functional electrical stimulation (FES). J. Electromyo. Kinesi. 2014, 24, 795–802.
[CrossRef] [PubMed]
69. Schalk, G.; McFarland, D.; Hinterberger, T.; Birbaumer, N.; Wolpaw, J.R. BCI2000: A General-Purpose
Brain-Computer Interface (BCI) System. IEEE Trans. Biomed. Eng. 2004, 51, 1034–1043. [CrossRef] [PubMed]
70. Delorme, A.; Makeig, S. EEGLAB: An open source toolbox for analysis of single-trial EEG dynamics including
independent component analysis. J. Neurosci. Methods 2004, 134, 9–21. [CrossRef]
71. Nolan, H.; Whelan, R.; Reilly, R.B. FASTER: Fully Automated Statistical Thresholding for EEG artifact
Rejection. J. Neurosci. Methods 2010, 192, 152–162. [CrossRef] [PubMed]
72. Schalk, G.; Mellinger, J. A Practical Guide to Brain—Computer Interfacing with BCI2000: General-Purpose Software
for Brain-Computer Interface Research, Data Acquisition, Stimulus Presentation, and Brain Monitoring; Springer
Science & Business Media: Berlin/Heidelberg, Germany, 2010.
73. Pfurtscheller, G.; Neuper, C. Future prospects of ERD / ERS in the context of brain-computer interface (BCI)
developments. Prog. Brain Res. 2006, 159, 433–437.
Brain Sci. 2020, 10, 512 26 of 27
74. Golub, M.D.; Chase, S.M.; Batista, A.P.; Yu, B.M. Brain–computer interfaces for dissecting cognitive processes
underlying sensorimotor control. Curr. Opin. Neurobiol. 2016, 37, 53–58. [CrossRef]
75. Nam, C.S.; Nijholt, A.; Lotte, F. Brain–Computer Interfaces Handbook: Technological and Theoretical Advances;
CRC Press: Boca Raton, FL, USA, 2018.
76. Pfurtscheller, G.; Lopes, F.H. Event-related EEG/MEG synchronization and desynchronization: Basic
principles. Clin. Neurophysiol. 1999, 110, 1842–1857. [CrossRef]
77. Novi, Q.; Guan, C.; Dat, T.H.; Xue, P. Sub-band Common Spatial Pattern (SBCSP) for Brain-Computer
Interface. In Proceedings of the 3rd International IEEE/EMBS Conference on Neural Engineering, Kohala
Coast, HI, USA, 2–5 May 2007; IEEE: Piscataway, NJ, USA, 2007; pp. 204–207.
78. Bashashati, A.; Fatourechi, M.; Ward, R.K.; Birch, G.E.E. A survey of signal processing algorithms in
brain–computer interfaces based on electrical brain signals. J. Neural Eng. 2007, 4, R32–R57. [CrossRef]
79. Polikar, R. Ensemble based systems in decision making. IEEE Circuits Syst. Mag. 2006, 6, 21–45. [CrossRef]
80. Kohavi, R. A study of cross-validation and bootstrap for accuracy estimation and model selection. IJCAI
1995, 14, 1137–1145.
81. Pfurtscheller, G.; Neuper, C. Motor imagery and direct brain-computer communication. Proc. IEEE 2001, 89,
1123–1134. [CrossRef]
82. Tonet, O.; Tecchio, F.; Sepulveda, F.; Citi, L.; Rossini, P.M.; Marinelli, M.; Tombini, M.; Laschi, C.;
Dario, P. Critical Review and Future Perspectives of Non-Invasive Brain-Machine Interfaces. European
Space Agency, the Advanced Concepts Team, Ariadna Final Report (05-6402). 2006. Available
online: https://www.esa.int/gsp/ACT/doc/ARI/ARI%20Study%20Report/ACT-RPT-BIO-ARI-056402-Non_
invasive_brain-machine_interfaces_-_Pisa_S%27Anna.pdf (accessed on 20 June 2020).
83. Ortner, R.; Allison, B.Z.; Korisek, G.; Gaggl, H.; Pfurtscheller, G. An SSVEP BCI to Control a Hand Orthosis
for Persons with Tetraplegia. IEEE Trans. Neural Syst. Rehabil. Eng. 2010, 19, 1–5. [CrossRef] [PubMed]
84. Nicolas-Alonso, L.F.; Gomez-Gil, J. Brain Computer Interfaces, a Review. Sensors 2012, 12, 1211–1279.
[CrossRef] [PubMed]
85. Vidaurre, C.; Sannelli, C.; Müller, K.-R.; Blankertz, B. Machine-Learning-Based Coadaptive Calibration for
Brain-Computer Interfaces. Neural Comput. 2011, 23, 791–816. [CrossRef]
86. Wolpaw, J.R.; McFarland, D.J. Control of a two-dimensional movement signal by a noninvasive
brain-computer interface in humans. Proc. Natl. Acad. Sci. USA 2004, 101, 17849–17854. [CrossRef]
87. Jung, T.P.; Makeig, S.; Humphries, C.; Lee, T.W.; McKeown, M.J.; Iragui, V.; Sejnowski, T.J. Removing
electroencephalographic artifacts by blind source separation. Psychophysiology 2000, 37, 163–178. [CrossRef]
88. Kuiken, T.A.; Li, G.; Lock, B.A.; Lipschutz, R.D.; Miller, L.A.; Stubblefield, K.A.; Englehart, K.B. Targeted
Muscle Reinnervation for Real-time Myoelectric Control of Multifunction Artificial Arms. JAMA 2009, 301,
619–628. [CrossRef]
89. Wodlinger, B.; Downey, J.E.; Tyler, E.C.; Schwartz, A.B.; Boninger, M.L.; Collinger, J.L. Ten-dimensional
anthropomorphic arm control in a human brain−machine interface: Difficulties, solutions, and limitations. J.
Neural Eng. 2014, 12, 16011. [CrossRef]
90. Wolpaw, J.R.; Ramoser, H.; McFarland, D.; Pfurtscheller, G. EEG-based communication: Improved accuracy
by response verification. IEEE Trans. Rehabil. Eng. 1998, 6, 326–333. [CrossRef]
91. Hart, S.G.; Staveland, L.E. Development of NASA-TLX (Task Load Index): Results of empirical and theoretical
research. Adv. Psychol. 1988, 52, 139–183.
92. Perelmouter, J.; Birbaumer, N. A binary spelling interface with random errors. IEEE Trans. Rehabil. Eng.
2000, 8, 227–232. [CrossRef] [PubMed]
93. Roset, S.A.; Gant, K.; Prasad, A.; Sanchez, J.C. An adaptive brain actuated system for augmenting rehabilitation.
Front. Neurosci. 2014, 8, 415. [CrossRef] [PubMed]
94. Fazel-Rezai, R.; Allison, B.Z.; Guger, C.; Sellers, E.W.; Kleih, S.C.; Kübler, A. P300 brain computer interface:
Current challenges and emerging trends. Front. Neuroeng. 2012, 5, 14. [CrossRef] [PubMed]
95. Rohm, M.; Schneiders, M.; Müller, C.; Kreilinger, A.; Kaiser, V.; Müller-Putz, G.R.; Rupp, R.R. Hybrid
brain–computer interfaces and hybrid neuroprostheses for restoration of upper limb functions in individuals
with high-level spinal cord injury. Artif. Intell. Med. 2013, 59, 133–142. [CrossRef] [PubMed]
96. Vuagnat, H.; Chantraine, A. Shoulder pain in hemiplegia revisited: Contribution of functional electrical
stimulation and other therapies. J. Rehabil. Med. 2003, 35, 49–56. [CrossRef]
Brain Sci. 2020, 10, 512 27 of 27
97. Blokland, Y.; Spyrou, L.; Thijssen, D.; Eijsvogels, T.M.; Colier, W.; Floor-Westerdijk, M.; Vlek, R.; Bruhn, J.;
Farquhar, J. Combined EEG-fNIRS Decoding of Motor Attempt and Imagery for Brain Switch Control:
An Offline Study in Patients With Tetraplegia. IEEE Trans. Neural Syst. Rehabil. Eng. 2013, 22, 222–229.
[CrossRef]
98. Huang, D.; Qian, K.; Fei, D.-Y.; Jia, W.; Chen, X.; Bai, O. Electroencephalography (EEG)-Based Brain–Computer
Interface (BCI): A 2-D Virtual Wheelchair Control Based on Event-Related Desynchronization/Synchronization
and State Control. IEEE Trans. Neural Syst. Rehabil. Eng. 2012, 20, 379–388. [CrossRef]
99. Yu, T.; Li, Y.; Long, J.; Gu, Z. Surfing the internet with a BCI mouse. J. Neural Eng. 2012, 9, 36012. [CrossRef]
100. Yu, T.; Li, Y.; Long, J.; Li, F. A Hybrid Brain-Computer Interface-Based Mail Client. Comput. Math. Methods
Med. 2013, 2013, 1–9. [CrossRef]
101. Kim, T.; Kim, S.; Lee, B. Effects of Action Observational Training Plus Brain-Computer Interface-Based
Functional Electrical Stimulation on Paretic Arm Motor Recovery in Patient with Stroke: A Randomized
Controlled Trial. Occup. Ther. Int. 2015, 23, 39–47. [CrossRef]
102. Looned, R.; Webb, J.; Xiao, Z.G.; Menon, C. Assisting drinking with an affordable BCI-controlled wearable
robot and electrical stimulation: A preliminary investigation. J. Neuroeng. Rehabil. 2014, 11, 51. [CrossRef]
[PubMed]
103. Toriyama, H.; Ushiba, J.; Ushiyama, J. Subjective Vividness of Kinesthetic Motor Imagery Is Associated with
the Similarity in Magnitude of Sensorimotor Event-Related Desynchronization Between Motor Execution
and Motor Imagery. Front. Hum. Neurosci. 2018, 12, 295. [CrossRef] [PubMed]
104. Keller, T.; Kuhn, A. Electrodes for transcutaneous (surface) electrical stimulation. J. Autom. Control. 2008, 18,
35–45. [CrossRef]
© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).